zrwangems/dsh-zrwang-plugin / ars/skills/deep-research
deep-research
Flagged学术研究前置:把模糊想法收敛成可研究的问题,并做文献综述、事实核查、系统综述(PRISMA)与三向文献扫描。Triggers: research, deep research, literature review, systematic review, meta-analysis, PRISMA, fact-check, guide my research | 研究, 文献综述, 系统性文献回顾, 事实核查, 引导我研究, 帮我确定研究问题。输出 RQ Brief、方法论蓝图、带核验的注释书目。
_shared/references/protected_hedging_phrases.md
View on GitHub> 输出语言(本 DSH 集成强制):所有面向用户的文字、稿件、摘要、评审与报告一律使用**简体中文**;英文术语与引文保持英文。文中的 `zh-TW` / `zh-tw-en` 只是上游语言代码与脚本参数,不代表要用繁体输出。用户明确要求其他语言时以用户为准。
>
> 多角色执行(评审团、十阶段流水线、角色扮演)必须使用 DSH 的 Agent Team 工具(`spawn_teammate`、`team_task_*`、`wait_agent`、`send_message`);当前会话看不到这些工具,说明用户没开 Agent Team,必须先提示用户开启,不得退化成单会话顺序扮演。
# Protected Hedging Phrases
**Spec:** v3.6.7 §7.1 — pattern protection reference for `report_compiler_agent` (abstract-only mode), Pattern C1.
**Audience:** Upstream calibration agents (academic-paper Stage 4, deep-research Phase 3 abstract handoff). The downstream `report_compiler_agent` reads this reference indirectly through the dispatch context that upstream calibration produces.
**Why this exists:** Live ARS pipeline runs surfaced a recurring failure where the abstract compiler, under hard word-count pressure (i.e., a fixed word budget set by the publisher), dropped epistemic hedges that the body of the paper depended on. The fix routes through a roster of protected hedging phrases that ride in the dispatch context to the compiler. A claim qualified by "may," "tentative," "preliminary," or "in this institutional context" in the body became unconditional in the abstract. The compression-driven drift created compression overclaim — the abstract no longer accurately represented the paper's epistemic stance, which is a publication-integrity failure.
The fix is to make hedging phrases **budget-protected**: upstream calibration marks specific phrases as protected, the protected list rides in the dispatch context to the abstract compiler, and the compiler honours the protection ahead of any other compression target.
---
## What counts as a protected hedging phrase
A phrase qualifies as protected when **dropping it changes the paper's truth-claim**, not merely its rhetorical register. Three categories cover most cases:
### 1. Epistemic hedges that bound the claim
Words and phrases that mark a claim as conditional, tentative, or under-evidenced. Removing them upgrades the claim to assertion.
**Examples:**
- "may," "might," "could" (modal hedges on causal or predictive claims)
- "tentative," "preliminary," "exploratory" (status hedges on findings)
- "suggests," "indicates," "is consistent with" (inferential hedges, distinct from "demonstrates," "proves," "establishes")
- "in our sample," "for this cohort," "under the conditions tested" (scope hedges)
**Why protected:** A reader who acts on an unconditional version of the claim takes on more risk than the evidence justifies. In academic abstracts this is overclaim; in policy abstracts this is misinformation.
### 2. Reflexivity / positionality markers
Phrases that disclose the author's relationship to the studied phenomenon. Removing them obscures the position from which the work was done.
**Examples:**
- "from a former evaluator's perspective" (positional)
- "as participants in the [year-range] reform" (involvement)
- "the authors served on the [committee] from [year] to [year]" (institutional role)
- "this analysis draws on the first author's experience as [role]" (experiential)
**Why protected:** Disclosure obligations under most journal policies (BMJ, Nature, social-science conventions) require positionality to ride in the abstract when it materially shapes interpretation. Stripping disclosure for word count violates the policy, not just the rhetorical norm.
### 3. Temporal disambiguation markers
Phrases that pin a claim or a positional statement to a specific time window. Removing them creates temporal ambiguity that downstream readers cannot resolve.
**Examples:**
- "between 2020 and 2024" (explicit year range)
- "during their term as [role] in [year]" (role-bounded period)
- "former [role]" (closed temporal status)
- "at the time of the [event]" — only protected when the event is named immediately adjacent; otherwise the deictic phrase is itself the failure mode (see `feedback_ars_phase2_phase3_downstream_agent_patterns.md`).
**Why protected:** Deictic temporal phrases ("during this period," "at the time," "currently") read differently when separated from their anchoring context. The abstract has no neighbouring context; an abstract-level deictic is unresolvable.
---
## Upstream calibration: how to mark phrases as protected
Upstream calibration runs at the end of paper drafting, after the body text is stable but before abstract compilation. The calibration agent walks the paper and identifies phrases in the three categories above, then emits a `protected_hedges` block in the dispatch payload to the abstract compiler.
### Dispatch block shape
```yaml
protected_hedges:
epistemic:
- phrase: "may"
anchored_at: "Section 4, claim about institutional adoption rate"
why: "claim is observational, not causal"
- phrase: "tentative"
anchored_at: "Section 5, conclusion about policy effect"
why: "n=12 sites, no comparison group"
reflexivity:
- phrase: "as a former evaluator on the [committee] between 2020 and 2024"
anchored_at: "Author note + Section 1"
why: "BMJ-equivalent disclosure obligation; informs how reader weights argument"
temporal:
- phrase: "between 2020 and 2024"
anchored_at: "Section 1, paragraph 2; Section 6, conclusion"
why: "abstract-level deictic ('at the time') would be unresolvable"
```
### Calibration rules
1. **Conservative inclusion.** When in doubt, include the phrase. Calibration cannot recover a hedge it did not list — the compiler treats every entry on the list as non-negotiable, so omitting a phrase removes that protection regardless of intent.
2. **Anchor every entry.** Each protected phrase must cite where in the paper it operates and one-line why. Without the anchor, the abstract compiler cannot judge replacement-vs-preservation when budget is tight.
3. **No duplicates.** One entry per phrase. The compiler counts protected phrases against the budget once.
4. **Calibration is mode-specific.** Deep-research INSIGHT abstracts and academic-paper journal abstracts have different convention (see `word_count_conventions.md`). Calibration runs once per target mode, not once per paper.
---
## Abstract compiler: how protection is honoured
The `report_compiler_agent` abstract-only mode treats protected hedges as **non-negotiable budget**. Concretely:
1. **Budget allocation order.** Protected hedges are reserved before any other content competes for budget, because they are non-negotiable.
- Step 1: Compute hard cap (whitespace-split, see `word_count_conventions.md`).
- Step 2: Reserve 3–5% buffer.
- Step 3: **Reserve words for protected hedges.** Sum the word count of every entry in `protected_hedges` (verbatim). This subtotal is non-negotiable budget.
- Step 4: Allocate the remaining words to required structural slots (research question, method, finding, implication).
- Step 5: Spend any words still remaining on rhetorical polish, transitions, secondary findings.
2. **Verbatim preservation rule.** A protected phrase rides verbatim into the abstract. The compiler does not paraphrase, compress, or substitute a synonym for a protected phrase. (A future v3.6.7+ extension may add an optional `approved_synonyms` field to the dispatch block; until then, verbatim only.)
3. **Cut order under budget pressure.** When the draft exceeds budget, cuts come from Step 5 first (rhetorical polish), then Step 4 (compress structural slots), **never** from Step 3 (protected hedges).
4. **Failure surface.** If the abstract cannot fit Steps 1–4 within hard cap, the compiler **reports the conflict** rather than dropping a protected hedge. The conflict goes back to upstream calibration to either renegotiate the protected list or the publisher's word cap.
---
## Where this protocol is enforced
The `report_compiler_agent` abstract-only mode prompt enforces the dispatch-context protocol and the budget allocation order. The authoritative protection clause lives in `../deep-research/agents/report_compiler_agent.md` under `PATTERN PROTECTION (v3.6.7)`. This file defines what counts as a protected hedge, the upstream calibration shape, and the compiler's budget allocation order; the agent prompt cites this file by path. Cross-model audit covers protected-hedge preservation under the report_compiler bundle's Section 4(f) check in `_shared/templates/codex_audit_multifile_template.md`.
---
## Cross-references
- `_shared/references/word_count_conventions.md` — whitespace-split standard, 3–5% buffer rule, publisher conventions.
- `_shared/templates/codex_audit_multifile_template.md` — audit dimensions including compression overclaim detection.
- `docs/design/2026-04-29-ars-v3.6.7-downstream-agent-pattern-protection-spec.md` §3.3 (C1) — pattern definition and provenance.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |